=Paper=
{{Paper
|id=None
|storemode=property
|title=Learning of Ontologies from the Web: the Analysis of Existent Approaches
|pdfUrl=https://ceur-ws.org/Vol-701/paper3.pdf
|volume=Vol-701
|dblpUrl=https://dblp.org/rec/conf/icdt/Omelayenko01
}}
==Learning of Ontologies from the Web: the Analysis of Existent Approaches==
https://ceur-ws.org/Vol-701/paper3.pdf
Learning of Ontologies for the Web: the Analysis of Existent
Approaches
Borys Omelayenko
Vrije Universiteit Amsterdam,
Division of Mathematics and Computer Science,
De Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands
Email: borys@cs.vu.nl, URL: www.cs.vu.nl/~borys
Abstract representation of the semantics of data accompanied
The next generation of the Web, called Semantic by domain theories (i.e. ontologies) will enable a
Web, has to improve the Web with semantic Knowledge Web that provides a qualitatively new
(ontological) page annotations to enable level of service. It will weave together a net linking
knowledge-level querying and searches. Manual incredibly large segments of human knowledge and
construction of these ontologies will require complement it with machine processability.
tremendous efforts that force future integration of This will require enrichment of the entire Web
machine learning with knowledge acquisition to with lots of ontologies that capture the domain
enable highly automated ontology learning. In the theories. Their manual construction will require
paper we present the state of the-art in the field of
enormous human efforts, thus ontology acquisition
ontology learning from the Web to see how it can
contribute to the task of semantic Web querying. becomes a bottleneck of the Semantic Web.
We consider three components of the query Recently ontologies have become a hot topic in the
processing system: natural language ontologies, areas of knowledge engineering, intelligent
domain ontologies and ontology instances. We information integration, knowledge management, and
discuss the requirements for machine learning electronic commerce [Fensel, 2000]. Ontologies are
algorithms to be applied for the learning of the knowledge bodies that provide a formal
ontologies of each type from the Web documents, representation of a shared conceptualization of a
and survey the existent ontology learning and other particular domain. Modern research focus lies in
closely related approaches. Web-based ontology representation languages based
on XML and RDF standards and further application
of ontologies on the Web (see [Decker et al., 2000]).
Introduction
Ontology learning (OL) is an emerging field aimed
Nowadays the Internet contains a huge collection of at assisting a knowledge engineer in ontology
data stored in billions of pages and it is used for the construction and semantic page annotation with the
worldwide exchange of information. The pages help of machine learning (ML) techniques.
represent mainly textual data and have no semantic In the next section of the paper we discuss the
annotation. Thus, query processing based mostly on general scheme for semantic querying of the Web
inefficient keyword-matching techniques becomes a with three ontological components required; the
bottleneck of the Web. subsequent sections discuss OL tasks and available
Tim Berners-Lee coined the vision of the next ML techniques. The survey section describes the
version of the Web, called Semantic Web [Berners- applications of ML techniques for the learning of
Lee&Fischetti, 1999], that would provide much different ontology types, and we conclude with
more automated services based on machine- comparison of the approaches.
processable semantics of data and heuristics that
make use of these metadata. The explicit
Semantic Querying of the Web
In: Proceedings of the International Workshop on Web Dynamics, In this section we discuss the general scheme for
held in conj. with the 8th International Conference on Database semantic querying of the Web, the types of
Theory (ICDT’01), London, UK, 3 January 2001
16
ontologies involved in query process, and basic ML frequently while the ontology of the catalogue will
algorithms available for learning the ontologies. remain the same).
The Semantic Web will require creation and
The General Scheme maintenance of a huge number of the ontologies of
The general scheme of the querying process is all three types, and the following ontology learning
presented in Figure 1. First, the user formulates the tasks will become important.
query in natural language. Then the query is
transformed into a formal query with the help of the Ontology Learning Tasks
natural language ontology and the domain ontology.
The Web pages are (possibly incomplete) instances Previous research in the area of ontology acquisition
of some domain ontologies, and they will contain proposed lots of guidelines for manual ontology
pieces of data semantically marked up according to development (see [Lopez, 1999] for an overview)
the underlying domain ontology. The query that organize the work of the knowledge engineer,
processor has to find the mapping between the but they pay no attention to the process of the
concepts of the initial query, the domain model used acquiring of the ontology by humans. The human
to expand the query, and the ontology instances on experts have to evolve the best knowledge
the Web. This mapping will be non-trivial and will acquisition process themselves from their past
require inference over domain ontologies. experience acquired by passing through numerous
case studies. Thus, we have to separate several tasks
Ontological Components in OL on our own:
There are a number of domains where ontologies Ontology creation from scratch by the knowledge
were successfully applied. The three ontologies that engineer. In this task ML assists the knowledge
are important for querying the Web (see Figure 1) engineer by suggesting the most important relations
are: in the field or checking and verifying the constructed
Natural Language Ontologies (NLO) that knowledge bases.
contain lexical relations between the language Ontology schema extraction from Web
concepts; they are large in size and do not require documents. In this task ML systems take the data
frequent updates. Usually they represent the and meta-knowledge (like a meta-ontology) as input
background knowledge of the system and are used to and generate the ready-to-use ontology as output
expand user’s queries. These ontologies belong to with the possible help of the knowledge engineer.
so-called ‘horizontal’ ontologies that try to capture Extraction of ontology instances populates given
all possible concepts, but they do not provide ontology schemas and extracts the instances of the
detailed description of each of the concepts. ontology presented in the Web documents. This task
Domain ontologies capture knowledge of one is similar to information extraction and page
particular domain, i.e. pharmacological ontology, or annotation and can apply the techniques developed in
printer ontology. These ontologies provide detailed these areas.
description of the domain concepts from a restricted Ontology integration and navigation deals with
domain (so-called ‘vertical’ontologies). Usually they reconstructing and navigating in large and possibly
are constructed manually but different learning machine-learned knowledge bases. For example, the
techniques can assist the (especially inexperienced) task can be to change the propositional-level
knowledge engineer. knowledge base of the machine learner into a first-
Ontology instances represent the main piece of order knowledge base.
knowledge presented in the future Semantic Web. As Ontology update task updates some parts of the
today’s Web is full of HTML documents of different ontology that are designed to be updated (like
layout, the future Web will be full of instances of formatting tags that have to track the changes made
different domain ontologies. The ontology instances in the page layout).
will serve as the Web pages and will contain the Ontology enrichment (or ontology tuning)
links to other instances (similar to the links to other includes automated modification of minor relations
Web pages). They can be generated automatically into existing ontology. This does not change major
and frequently updated (i.e. a company profile from concepts and structures but makes the ontology more
the Yellow Pages catalogue will be updated precise. Unlike ontology update, this task deals with
17
the relations that are not specially designed to be Bayesian learning is mostly represented by Naive
updated. Bayes classifier. It is based on the Bayes theorem
The first three tasks relate to ontology acquisition and generates probabilistic attribute-value rules
tasks in knowledge engineering, and the next three to based on the assumption of conditional independence
ontology maintenance tasks. In this paper we do not between the attributes of the training instances.
consider ontology integration and ontology update First-order logic rules learning induces the rules
tasks. that contain variables, called first-order Horn
clauses. The algorithms usually belong to the FOIL
family of algorithms and perform general-to-specific
Machine Learning Techniques hill-climbing search for the rules that cover all
The main requirement for ontology representation is available positive training instances. With each
that ontologies must be symbolic, human-readable iteration it adds one more literal to specialize the rule
and understandable. This forces us to deal only with until it avoids all negative instances.
symbolic learning algorithms that make Clustering algorithms group the instances
generalizations and skip other methods, like neural together based on the similarity or distance measures
networks, genetic algorithms and the family of 'lazy between a pair of instances defined in terms of their
learners' (see [Mitchell, 1997] for an introduction to attribute values. Various search strategies can guide
ML and the algorithms mentioned below). The the clustering process. Iterative application of the
foreseen potentially applicable ML algorithms algorithm will produce hierarchical structures of the
include: concepts.
Propositional rule learning algorithms that learn The knowledge bases built by ML techniques
association rules, or other attribute-value rules. The substantially differ from the knowledge bases that
algorithms are generally based on a greedy search of we call ontologies. The differences are inspired by
the attribute-value tests that can be added to the rule the fact that ontologies are constructed to be used by
preserving its consistency with the set of training humans, while ML knowledge bases are only used
instances. Decision tree learning algorithms, mostly automatically. This leads to several differences listed
represented by the C4.5 algorithm and its in Table 1.
modifications, are used quite often to produce high- To enable automatic OL we must adapt ML
quality propositional-level rules. The algorithm uses techniques so that they can automatically construct
statistical heuristics over the training instances, like ontologies with the properties of manually
entropy, that guide hill-climbing search of the constructed ontologies. Thus, OL techniques have to
decision trees. Learned decision trees are equivalent possess the following properties, which we trace in
to the sets of propositional-level classification rules the survey:
that are conjunctions of attribute-value tests. - ability to interact with a human to acquire his
http://www.cs…
Formal Web pages:
http://www.cs…
Natural Semantic ontology
Language Query to the
Web pages:
http://www.cs…
Query instances
ontology
Web Web pages:
instances
ontology
instances
Domain
Domain
Ontologies
Natural Domain
Ontologies
Language Ontologies Instance-of
Ontology links
Figure 1. Semantic querying of the Web
18
knowledge and to assist him; this requires the actions. Concept features are usually represented
readability of internal and external results of the by adjectives or adjective nouns (like ‘strong-
learner; strength’). Thus the ontology can be represented by
- ability to use complex modelling primitives; frames with a limited structure.
- ability to deal with complex solution space, NLOs define the first and basic interpretation of
including composed solutions. user’s query, and they must link the query to specific
Each ontology type has special requirements for terminology and specific domain ontology. General
ML algorithms applied for learning these types of language knowledge contained in a general-purpose
ontologies. NLO like WordNet [Fellbaum, 1998] is not
sufficient for such a purpose. In order to achieve
Table 1. Manual and machine representations
this, lots of research efforts have been focused on
Machine-learned Manually constructed NLO enrichment. NLO enrichment from domain
knowledge bases ontologies texts is a suitable task for ML algorithms, because it
Modelling primitives provides a good set of training data for the learner
Simple and limited. For Rich set of modelling (the corpus).
example, decision tree primitives (frames,
learning algorithms gene- subclass relation, rules
NLOs do not require either frequent or automatic
rate the rules in the form of with rich set of updates. They are updated from time to time with
conjunctions over attribute- operations, functions, intensive cooperation from a human, thus ML
value tests. etc.). algorithms for NLO learning are not required to be
Knowledge base structure fast.
Flat and homogeneous. Hierarchical, consists of Domain ontologies use the whole set of modelling
various components with primitives, like (multiple) inheritance, numerous
subclass-of, part-of and slots and relations, etc. They are complex in
other relations. structure and are usually constructed manually.
Tasks Domain ontology learning concentrates on
Classification and Classification task discovering statistically valid patterns in the data in
clusterization that map the requires mapping of
order to suggest them to the knowledge engineer who
objects described by the objects into a tree of
attribute-value pairs into a structured classes. It can guides the ontology acquisition process. In future we
limited and unstructured set require construction of would like to see an ML system that guides this
of class or cluster labels. class descriptions instead process and asks the human to validate pieces of the
of selection. constructed ontology.
Problem-solving methods ML will be used to predict the changes made by
Very primitive, based on Complicated, require the human to reduce the number of interactions. The
simple search strategies, inference over a input of this learner will consist of the ontology
like hill-climbing in knowledge base with a being constructed, human suggestions and domain
decision tree learning. rich structure, often knowledge.
domain-specific and Domain ontologies require more frequent updates
application-specific. than NLOs (just as new technical objects appear
Solution space
The non-extensible, fixed Extensible set of
before the community has agreed about the
set of class labels. primitive and compound surrounding terminology), their updates are done
solutions. manually and ML algorithms that assist this process
Readability of the knowledge bases to a human are also not required to be fast.
Not required. They can be Required. They may be Ontology instances are contained in the Web
used only automatically and (at least potentially) used pages marked up with the concepts of the underlying
only in special domains. by humans. domain ontology with information extraction or
annotation rules. The instances will require more
NLO contain hierarchical clustering of the frequent updates than domain ontologies or NLOs
language concepts (words and their senses). The set (i.e. a company profile in a catalogue will be
of relations (slots) used in the representation is updated faster than the ontology of a company
limited. The main relations between the concepts are: catalogue).
‘synonyms’, ‘antonyms’, ‘is-a’, ‘part-of’. The verbs
can contain several additional relations to describe
19
word ‘waiter’ has two senses: the waiter in the
restaurant (related words: waiter–restaurant,
The Survey menu, dinner); and a person who waits (related
This section presents the survey of existing words: waiter–station, airport, hospital). The
techniques related to the learning and enriching of system queries the Web for the documents related to
the NLO from the Web, Web-based support for each concept from the WordNet and then builds a
domain ontology construction, and extraction of list of words associated with the topic. The lists are
ontology instances. These approaches cover various called topic signatures and contain the weight (called
issues in the field and show different applications of strength) of each word. The documents are retrieved
ML techniques. by querying the Web with the AltaVista search
engine by asking for the documents that contain the
Learning of NLO words related to a particular sense and do not
contain the words related to the other senses of the
Lots of conceptual clustering methods can be used word. A typical query may look something like
for ontology construction but no methodology or tool ‘waiter AND (restaurant OR menu) AND NOT
has been developed to support the elaboration of (station OR airport)’ to get the documents that
conceptual clustering methods that build task- correspond to the ‘waiter, server’ concept.
specific ontologies. The Mo'K tool [Bisson et al., NLOs, like EuroWordNet or WordNet, help in the
2000] supports development of conceptual clustering understanding of natural language queries and in
methods for ontology building. The paper focuses on bringing semantics to the Web. But in specific
elaboration of the clustering methods to perform domains general language knowledge becomes
human-assisted learning of conceptual hierarchies insufficient and that requires creation of domain-
from corpora. The input for the clustering methods is specific NLOs. Early attempts to create such domain
represented by the classes (nouns) and their ontologies to perform information extraction from
attributes (grammatical relations) received after texts failed because the experts used to create the
syntactical analysis of the corpora, which are in turn ontologies with lots of a priori information that was
characterized by the frequency with which they not reflected in the texts. The paper
occur in the corpora. [Faure&Poibeau, 2000] suggests improving NLO by
The algorithm uses bottom-up clustering to group unsupervised domain-specific clustering of texts
'similar' objects to create the classes and to from corpora. The system Asium described in the
subsequently group 'similar' classes to form the paper cooperatively learns semantic knowledge from
hierarchy. The user may adjust several parameters of texts which are syntactically parsed, without
the process to improve performance: select input previous manual processing. It uses the syntactic
examples and their attributes, level of pruning, and parser Sylex to generate the syntactical structure of
distance evaluation functions. The paper presents an the texts. Asium uses only head nouns of
experimental study that illustrates how learning complements and links to verbs and performs
quality depends on the different combinations of bottom-up breadth-first conceptual clustering of the
parameters. corpora to form the concepts of ontology level. On
While the system allows the user to tune its each level it allows the expert to validate and/or
parameters, it performs no interactions during label the concepts. The system generalizes the
clustering. It builds the hierarchy of the frames that concepts that occur in the same role in the texts and
contain lexical knowledge about the concepts. The uses generalized concepts to represent the verbs.
input corpora can be naturally found on the Web, Thus, state of the art in NLO learning looks
and the next paper presents a way of integrating quite optimistic: not only does a stable general-
NLO enrichment with the Web search of the relevant purpose NLO exist but so do techniques for
texts. automatically or semiautomatically constructing
The system [Agirre et al., 2000] exploits the text and enriching domain-specific NLO.
from the Web to enrich the concepts in the WordNet
[Fellbaum, 1998] ontology. The proposed method
Learning of Domain Ontologies
constructs lists of topically related words for each
concept in the WordNet, where each word sense has Domain-specific NLO significantly improves
one associated list of related words. For example, the semantic Web querying but in specific domains
general language knowledge becomes insufficient
20
and query processing requires special domain in terms of the common understanding of the
ontologies. domain, i.e. in the terms of the domain ontology.
The paper [Maedche&Staab, 2000] presents an The system for ontology-based induction of high-
algorithm for semiautomatic ontology learning from level classification rules [Taylor et al., 1997] goes
texts. The learner uses a kind of algorithm for further and uses ontologies not only to explain the
discovering generalized association rules. The input discovered rules for a user, but also to guide learning
data for the learner is a set of transactions, each of algorithms. The algorithm consequently generates
which consists of a set of items that appear together queries for an external learner ParkaDB, that uses
in the transaction. The algorithm extracts association the domain ontology and the input data to check
rules represented by sets of items that occur together consistency of the query, and consistent queries
sufficiently often and presents the rules to the become classification rules. The query generation
knowledge engineer. For example, shopping process continues until the set of queries covers the
transactions may include the items purchased whole data set. Currently the domain ontologies used
together. The association rule may say that ‘snacks there are restricted to simple concept hierarchies
are purchased together with drinks’ rather than where each attribute has its own hierarchy of
‘crisps are purchased with beer’. The algorithm uses concepts. On the bottom level the hierarchy contains
two parameters: support and confidence for a rule. attribute values present in the data, the next level
Support is the percentage of transactions that contains a generalization about these attribute
contain all the items mentioned in the rule, and values. This forms one-dimensional concepts, and a
confidence for the rule X? Y is conditional domain ontology of a very specialized type.
percentage of transactions where Y is seen, given The approach uses a knowledge-base system and
that X also appeared in the transaction. The ontology its inference engine to validate classification rules. It
learner [Maedche&Staab, 2000] applies this method generates the rules in terms of the underlying
straightforwardly for ontology learning from texts to ontology, where the ontology still has a very
support the knowledge engineer in the ontology restricted type.
acquisition environment. The paper [Webb, Wells, Zheng, 1999]
The main problem in applying ML algorithms for experimentally demonstrates how the integration of
OL is that the knowledge bases constructed by the machine learning techniques with knowledge
ML algorithms have a flat homogeneous structure, acquisition from experts can both improve the
and very often have prepositional level accuracy of the developed domain ontology and
representation (see Table 1). Thus several efforts reduce development time. The paper analyses three
focus on improving ML algorithms in terms of types of knowledge acquisition system: the systems
ability to work with complicated structures. for manual knowledge acquisition from experts, ML
The first step in applying ML techniques to systems and the integrated systems built for two
discover hierarchical relations between textually domains. The knowledge bases were developed by
described classes is taken with the help of Ripple- experienced computer users who were novices in
Down Rules [Suryanto&Compton, 2000]. The knowledge engineering.
authors start with the discovery of the class relations The knowledge representation scheme was
between classification rules. Three basic relations restricted to flat attribute-value classification rules
are considered: intersection (called subsumption in and the knowledge base was restricted to a set of
marginal cases) of classes, mutual-exclusivity, and production rules. The rationale behind this
similarity. For each possible relation they define a restriction was based on the difficulties that novice
measure to evaluate the degree of subsumption, users experience when working with first-order
mutual exclusivity, and similarity between the representations. The ML system used the C4.5
classes. For input, the measures use the attributes of decision tree learning algorithm to support the
the rules that lead to the classes. After the measures knowledge engineer and to construct the knowledge
between all classes have been discovered, simple bases automatically.
techniques can be used to create the hierarchical The use of machine learning with knowledge
(taxonomic) relations between the classes. acquisition by experts led to the production of more
Knowledge extraction from the Web (data mining accurate rules in significantly less time than
from the Web) uses domain ontologies to represent knowledge acquisition alone (up to eight times less).
the extracted knowledge to the user of the knowledge The complexity of the constructed knowledge bases
21
was mostly the same for all systems. The In a classical setting the algorithm C4.5 will take
questionnaire presented in the paper showed that the the instances described by attribute-value pairs and
users found the ML facilities useful and thought that produce a tree with nodes that are attribute-value
they made the knowledge acquisition process easier. tests. The authors propose replacing the attribute-
Future prospects for research listed in [Webb, value dictionary with a more expressive one that
Wells, Zheng, 1999] were to lead to ‘a more consists of simple data types, tuples, sets, and
ambitious extension of this type of study that would graphs. The method [Bowers et al., 2000] uses a
examine larger scale tasks that included the modified C4.5 learner to generate a classification
formulation of appropriate ontologies’. tree that consists of tests on these structures, as
Learning of the domain ontologies is far less opposed to attribute value tests in a classical setting.
developed than NLO improvement. The acquisition Experiments showed that on the data sets with
of the domain ontologies is still guided by a structured instances the performance of this
human knowledge engineer, and automated algorithm is comparable to standard C4.5 but task-
learning techniques play a minor role in oriented modifications of C4.5 perform much better.
knowledge acquisition. They have to find The system CRYSTAL [Soderland et al., 1995]
statistically valid dependencies in the domain texts extends the ideas of the previous system AutoSlog,
and suggest them to the knowledge engineer. which showed great performance increase (about
200 times better than the manual system) on a
Learning of Ontology Instances creation of concept node definitions for a terrorism
In this subsection we survey several methods for domain. It uses an even richer set of modelling
learning of the ontology instances. primitives and creates the text extraction and mark-
The traditional propositional-level ML approach up rules, with a given domain model as input, by
represents knowledge about the individuals as a list generalizing semantic mark-up of the manually
of attributes, with each individual being represented marked-up training corpora. Manually created mark-
by a set of attribute-value pairs. The structure of up is automatically converted into a set of case
ontology instances is too rich to be adequately frames called ‘concept nodes’ using a dictionary of
captured by such a representation. The paper rules that can be present in the concept node. The
[Bowers et al., 2000] uses a typed, higher-order concept nodes represent the ontology instances and
logic to represent the knowledge about the the domain-specific dictionary of rules defines the
individuals. list of allowable slots in the ontology instance.
Table 2. Comparison of the ontology learning approaches
Type OL Task ML Modifications of ML techniques
technique
First-Order Rule learn.
Approach Human Complex Complex solution
Propositional learn.
Domain Ontologies
Ontology Instances
Instance Extraction
Bayesian learning
interaction modelling space
primitives
Enrichment
Clustering
Extraction
Creation
NLO
[Bisson et al., 2000] X X X Partial No Concept hierarchy
[Faure&Poibeau, 2000] X X X Yes Simplified frames Simplified frames
[Agirre et al., 2000] X X X No No No
[Junker et al., 1999] X X X No Several predicates No
[Craven et al., 2000] X X X X No No No
[Bowers et al., 2000] X X X No Yes, rich structure Yes, rich structure
[Taylor et al., 1997] X X X No Yes, but restricted No
[Webb, Wells, Zheng, 1999] X X X Yes No No
[Soderland et al., 1995] X X X X X No Yes Yes
[Maedche&Staab, 2000] X X X No No No
22
After formalizing the instance level of the position) and three predicates governing these types
hierarchy, CRYSTAL performs a search-based for treating text categorization rules as logical
generalization of the concept nodes. A pair of nodes programs and applying first-order rule learning
is generalized by creating a parent class with the algorithms. The rules learned are derived from five
attributes that both classes have in common. basic constructs of a logical pattern language used in
The knowledge representation language for the the framework to define the ontologies. The learned
concept nodes is very expressive, which leads to an rules are directly exploited in automated annotation
enormous branching factor for the search performed of the documents to become the ontology instances.
during the generalization. The system stores the The task of learning of the ontology instances
concept nodes in a way that best suits the distance fits nicely into an ML framework, and there are
measure function, and therefore performs reasonably several successful applications of ML algorithms
efficiently. Experiments on a medical domain for this. But these applications are either strictly
showed that the number of positive training instances dependent on the domain ontology or populate
required for a good recall was limited; after between the mark-up without relating to any domain
1 and 2 thousand, recall measure no longer grows theory. A general-purpose technique for
significantly. extracting ontology instances from texts given the
The system performs two stages necessary for OL: domain ontology as input has still not been
it formalizes ontology instances from the text and developed.
generates a concept hierarchy from these instances.
A systematic study of the extraction of ontology
instances from the Web documents was carried out Conclusions
in the project Web-KB [Craven et al., 2000]. In their The above case study is summarized in Table 2. The
paper the authors used the ontology of an academic first column specifies the approach; the next
web-site to populate it with actual instances and columns represent the ontological component of the
relations from CS departments’ web sites. The paper Web query system, the OL tasks, and the relevant
targets three learning tasks: ML technique respectively. The last three columns
(1) recognizing class instances from the hypertext describe the degree to which the system interacts
documents guided by the ontology; with the user and the properties of the knowledge
(2) recognizing relation instances from the chains representation scheme.
of hyperlinks; From the table we see that a number of systems
(3) recognizing class and relations instances from related to the natural-language domain deal with
the pieces of hypertext. domain-specific tuning and enrichment of the NLOs
The tasks are dealt with using two supervised with various clustering techniques.
learning approaches: Naive Bayes algorithm and Learning of the domain ontologies is done by now
first-order rule learner (modified FOIL). only on a propositional level, and first-order
The system automatically creates mapping representations are used only in the extraction of
between the manually constructed domain ontology ontology instances (see Table 2).
and the Web pages by generalizing from the training There are several approaches in the field of
instances. The system performance was surprisingly domain ontology extraction, but the systems used
good for the restricted domain of a CS website there are the variants of propositional-level ML
where it was tested. algorithms.
Major ML techniques applied for text Each OL paper modifies the applied ML algorithm
categorization performed to some degree of to handle human interaction, complex modelling
effectiveness [Junker et al., 1999], but beyond that, primitives or complex solution space together. Only
effectiveness appeared difficult to attain and was one paper [Faure&Poibeau, 2000] makes all three
only possible in a small number of isolated cases modifications of the ML algorithm for NLO
with substantial heuristic modification of the learning, as also shown in the table.
learners. This shows the need for combining these The research in OL goes mostly in the way of
modifications in a single framework based on first- straightforward application of the ML algorithms.
order rule learning. This was a successful strategy for beginning, but we
The paper [Junker et al., 1999] defines three basic would need substantial modifications of the ML
types (one for text, one for word, and one for text algorithms for OL tasks.
23
Workshop on Description Logics (DL2000),
Aachen, Germany, August, 2000.
Acknowledgements
The author would like to thank Dieter Fensel for [Evett, 1994] M. Evett. PARKA: A System for
helpful discussions and comments, and Heiner Massively Parallel Knowledge Representation.
Stuckenschmidt and four anonymous reviewers for Computer Science Department, University of
their comments. Maryland at College Park, 1994.
[Faure&Poibeau, 2000] D. Faure, T. Poibeau:
References First experiments of using semantic knowledge
learned by ASIUM for information extraction task
using INTEX. In: S. Staab, A. Maedche, C.
[Agirre et al., 2000] E. Agirre, O. Ansa, E. Hovy, Nedellec, P. Wiemer-Hastings (eds.), Proceedings
D. Martinez: Enriching very large ontologies using of the Workshop on Ontology Learning, 14th
the WWW. In: S. Staab, A. Maedche, C. Nedellec, European Conference on Artificial Intelligence
P. Wiemer-Hastings (eds.), Proceedings of the ECAI'00, Berlin, Germany, August 20-25, 2000.
Workshop on Ontology Learning, 14th European
Conference on Artificial Intelligence ECAI'00, [Fellbaum, 1998] C. Fellbaum. WordNet: An
Berlin, Germany, August 20-25, 2000. Electronic Lexical Database. The MIT Press,
1998.
[Berners-Lee&Fischetti, 1999] T. Berners-Lee, M.
Fischetti. Weaving the Web. Harper, San [Fensel, 2000] D. Fensel. Ontologies: Silver Bullet
Francisco, 1999. for Knowledge Management and Electronic
Commerce. Springer-Verlag, Berlin, 2000.
[Bisson et al., 2000] G. Bisson, C. Nedellec, D.
Canamero: Designing Clustering Methods for [Junker et al., 1999] M. Junker, M. Sintek, M.
Ontology Building - The Mo'K Workbench. In: S. Rinck: Learning for Text Categorization and
Staab, A. Maedche, C. Nedellec, P. Wiemer- Information Extraction with ILP. In: J. Cussens
Hastings (eds.), Proceedings of the Workshop on (eds.), Proceedings of the 1st Workshop on
Ontology Learning, 14th European Conference on Learning Language in Logic, Bled, Slovenia, June
Artificial Intelligence ECAI'00, Berlin, Germany, 1999, 1999, pp. 84-93.
August 20-25, 2000.
[Lopez, 1999] F. Lopez: Overview of
[Bowers et al., 2000] A. Bowers, C. Giraud- Methodologies for Building Ontologies. In:
Carrier, J. Lloyd: Classification of Individuals Proceedings of the IJCAI-99 workshop on
with Complex Structure. In: Proceedings of the Ontologies and Problem-Solving Methods,
Seventeenth International Conference on Machine Stockholm, Sweden, August 2, 1999.
Learning (ICML'2000), Stanford, US, June 29-July
2, 2000, pp. 81-88. [Maedche&Staab, 2000] A. Maedche, S. Staab:
Semi-automatic Engineering of Ontologies from
[Craven et al., 2000] M. Craven, D. DiPasquo, D. Text. In: Proceedings of the Twelfth International
Freitag, A. McCallum, T. Mitchell, K. Nigam, S. Conference on Software Engineering and
Slattery: Learning to construct knowledge bases Knowledge Engineering, Chicago, 2000.
from the World Wide Web. Artificial Intelligence,
118: 69-113, 2000. [Mitchell, 1997] T. Mitchell. Machine Learning.
McGraw Hill, 1997.
[Decker et al., 2000] S. Decker, D. Fensel, F. van
Harmelen, I. Horrocks, S. Melnik, M. Klein, J. [Soderland et al., 1995] S. Soderland, D. Fisher, J.
Broekstra: Knowledge Representation on the Web. Aseltine, W. Lehnert: Issues in Inductive Learning
In: Proceedings of the 2000 International of Domain-Specific Text Extraction Rules. In:
Proceedings of the Workshop on New Approaches
24
to Learning for Natural Language Processing at [Taylor et al., 1997] M. Taylor, K. Stoffel, J.
IJCAI'95, Montreal, Quebec, Canada, 1995.
Hendler: Ontology-based Induction of High Level
[Studer et al., 1998] R. Studer, R. Benjamins, D. Classification Rules. In: Proceedings of the
Fensel: Knowledge Engineering: Principles and SIGMOD Data Mining and Knowledge Discovery
methods. Data and Knowledge Engineering, 25: workshop, Tuscon, Arizona, 1997.
161-197, 1998. [Webb, Wells, Zheng, 1999] G. Webb, J. Wells, Z.
[Suryanto&Compton, 2000] H. Suryanto, P. Zheng: An Experimental Evaluation of Integrating
Compton: Learning Classification taxonomies from Machine Learning with Knowledge Acquisition.
a classification knowledge based system. In: S. Machine Learning, 31(1): 5-23, 1999.
Staab, A. Maedche, C. Nedellec, P. Wiemer-
Hastings (eds.), Proceedings of the Workshop on
Ontology Learning, 14th European Conference on
Artificial Intelligence ECAI'00, Berlin, Germany,
August 20-25, 2000.
25